Load EmbeddingGemma 2 text-only, and do not use float16

The full checkpoint is 740M parameters because the vision encoder is 170M and the audio encoder is 300M. A text search index does not need either. The card shows how to leave them unloaded.

This is the memory path. The search prompts are in the local-search recipe. Do both if you are indexing documents and you are short on RAM.

From the card's selective-encoder table. Text only is vision_config and audio_config set to None, effective size 270M.

from sentence_transformers import SentenceTransformer

model = SentenceTransformer(
    "google/embeddinggemma-2",
    config_kwargs={"vision_config": None, "audio_config": None},
)

Other rows on that table, if you need them later:

• Text and image, audio off: {"audio_config": None} — 440M
• Text and audio, vision off: {"vision_config": None} — 570M
• Full multimodal: {} — 740M

The card says the omit syntax differs by library. This block is the Sentence Transformers form. Do not paste it into a different loader and assume the keys match.

The card says run inference in bfloat16 or float32. Do not use float16. The activation range exceeds float16, and the failure mode is NaN or a silently worse vector, not an exception.

Use float32 on most CPUs. Use bfloat16 where the runtime actually supports it. The card's Sentence Transformers example sets torch_dtype through model_kwargs. Match that example on the card if you need the exact check. Do not force float16 to "save memory."

Matryoshka dimensions on the card are 768, 512, 256, and 128. Shorter vectors need a re-normalize. Slicing a unit vector does not leave a unit vector. The card says skipping that step hurts ranking and still returns a plausible score, so you will not see an error.

Queries and documents have to share a dimension. A 768-d query against a 128-d corpus is not a search.

query_emb = model.encode(
    query,
    prompt_name="SearchQuery",
    truncate_dim=256,
    normalize_embeddings=True,
)

The card's truncation table says 256-d stays close to 768-d on the reported means. 128-d drops harder, especially on multimodal benches. Validate 128-d on your own text set before you shrink a production index to it.

• Model card, selective encoders, precision, and truncation: https://huggingface.co/google/embeddinggemma-2